v0.5.20

Reverted

• groq: Reverted 7 non-coding models incorrectly added in v0.5.19:
  • whisper-large-v3, whisper-large-v3-turbo (audio transcription — not coding)
  • canopylabs/orpheus-arabic-saudi, canopylabs/orpheus-v1-english (speech — not coding)
  • meta-llama/llama-prompt-guard-2-22m, meta-llama/llama-prompt-guard-2-86m (security guardrails — not coding)
  • openai/gpt-oss-safeguard-20b (safety — not coding)
  • groq remains at 8 coding models; total catalog back to 154 models.

Kept (from v0.5.19 audit — all valid coding changes)

• nvidia: Added nvidia/nemotron-3-ultra-550b-a55b (S+, 1M context, NVIDIA NIM)
• nvidia: Corrected context windows:
  • deepseek-ai/deepseek-v4-pro: 128k → 1M
  • deepseek-ai/deepseek-v4-flash: 128k → 1M
  • mistralai/mistral-small-4-119b-2603: 128k → 256k
• nvidia: Removed deprecated model z-ai/glm5 (replaced by z-ai/glm-5.1)

Notes

• This catalog focuses exclusively on coding LLMs — audio, speech, guardrail, and safety models are out of scope.
• Model counts: nvidia 27 (was 27 → no net change after removing glm5 and adding nemotron-ultra), groq back to 8.
• audit_state.json fingerprints corrected for groq.